The trial-and-error approach for calibrating the meso-parameters of the parallel bond model is cumbersome and time-consuming

and it fails to quantitatively assess the correlation between cracks from numerical simulations and laboratory tests. The meso-parameter ranges of the parallel bond model over the past decade were summarized

and the crack fractal dimensions post-failure in both numerical simulations and laboratory tests were calculated using the box-counting method. Based on this

a neural network model was developed using four macroscopic parameters

including elastic modulus

Poisson’s ratio

peak strength and crack fractal dimension

as inputs

and six mesoscopic parameters such as bond effective modulus

ratio of normal stiffness to shear stiffness

cohesion

friction angle

tensile strength and friction coefficient

as outputs. The calibration effects of the parallel bond model with and without considering crack fractal dimension were compared. The research indicates that: 1) The developed neural network model exhibits a good convergence rate

prediction accuracy

and generalization ability

with an error of approximately 3.34% between the test set output and the expected values. 2) Incorporating crack fractal dimension into the neural network model results in errors of less than 3.00% for macroscopic parameters like elastic modulus

peak strength

and Poisson’s ratio between numerical and laboratory tests

outperforming the calibration results without considering crack fractal dimension. 3) This approach quantitatively ensures the consistency of crack irregularity between numerical and laboratory tests. The calibrated results can partially correct the existing neural network model’s calibration

offering new insights for enhancing the calibration of the meso-parameters of the parallel bond model.